AI Model Successfully Monitors Keystroke Operations with an Accuracy Rate of Up to 95%

2023-08-08

Since the existence of personal computers, security experts have been urging people to use strong passwords. Even this reasonable and still underestimated security measure may have lost some of its usefulness now. Researchers in the UK have proven that it is easy to steal data simply by listening to someone's typing sound.



In the past, we have seen researchers successfully steal data in unusual ways, such as detecting the waste heat of computers. The clattering and clicking sounds of keyboards may sound like random noise, but with enough data, an artificial intelligence can be trained to understand it. And it can be surprisingly effective.


The researchers used the built-in keyboard of a MacBook Pro for the experiment, but theoretically any keyboard can be used - even those with almost no noise. They pressed each key on the keyboard 25 times and recorded the sound using an iPhone microphone placed 17 centimeters away.


These sounds were processed and enhanced to generate spectrograms, which were then input into an AI model called CoAtNet. This is a hybrid neural network that combines convolutional artificial intelligence with self-attention, enabling it to classify spectrograms from relatively small datasets. Therefore, this AI cannot read your password just from general keyboard sounds, it needs data from known sources. For example, mass-produced keyboards from one of the most popular laptop models.



After training this AI, the researchers reported that it could decode new keystroke sounds recorded by an iPhone with an astonishing accuracy of 95%. The study also included audio data recorded during video chats, with a slight decrease in accuracy. In Zoom calls, the keystroke accuracy of the AI reached 93%, while on Skype it was 92%. Therefore, forgetting to mute while typing during a video chat is not only annoying but also insecure.


The researchers proposed several possible defense methods to prevent such espionage activities. You can use a white noise machine to mask your keystroke sounds, or alternatively, applications on your computer can insert additional keystroke sounds to confuse the algorithm. You can also use a different keyboard or change your typing style. Anything that changes the quality of the waveform on which this method relies may help protect you. Customizing a mechanical keyboard with replaceable switches can also achieve the same effect, as suggested by Bleeping Computer. Using biometric recognition and password managers can completely prevent you from entering sensitive characters.


This research is currently a preprint, so it has not been peer-reviewed yet. We will wait and see if the team's research results are easily replicable. As interest in artificial intelligence grows, we may need to rethink how to protect data security.

RECENT AI NEWS

RECENT AI TOOLS